Rapid dataset generation methods for stacked construction solid waste based on machine vision and deep learning.

Area of Science:

• Computer Vision
• Environmental Science
• Robotics

Background:

• Urbanization generates substantial construction solid waste, necessitating efficient sorting solutions.
• Machine vision algorithms offer faster, more stable solid waste detection than traditional methods.
• Accurate machine vision requires large datasets, but field data is scarce and manual annotation is costly.

Purpose of the Study:

• To develop rapid, automatic methods for generating and annotating construction solid waste datasets.
• To improve the accuracy and efficiency of machine vision-based solid waste sorting.
• To address the limitations of scarce field data and high manual annotation costs.

Main Methods:

• An acquisition and detection platform was built for automatic RGB-D image collection and instance labeling.
• A rapid-generation method for synthetic construction solid waste datasets using distribution points and data augmentation.
• Two automatic annotation methods for real datasets: semi-supervised self-training and RGB-D fusion edge detection.

Main Results:

• The generated dataset achieved a 95.98 F1-score in simple conditions, surpassing manual labeling (94.81).
• In complex conditions, the rapid generation method reached a 97.74 F1-score, significantly outperforming manual labeling (85.97).
• Datasets generated and annotated using proposed methods yielded superior model training results.

Conclusions:

• The proposed rapid dataset generation and automatic annotation methods are effective for construction solid waste.
• These automated approaches overcome limitations of manual annotation and scarce field data.
• The developed methods enhance the performance of machine vision algorithms for solid waste sorting.